CycleGAN Variants for Industrial Defect Data Augmentation

International Campus, Zhejiang University, 718 Haizhou East Road, Haining, China

^* Corresponding author: xiaoyu.22@intl.zju.edu.cn

Abstract

Industrial visual inspection is constrained by scarce labeled defect samples and complex surface patterns in bearings, steel, and ICs, significantly hindering deep learning detection models. This review conducts a systematic comparative analysis of CycleGAN variants tailored for industrial defect data augmentation, focusing on how they address generic CycleGAN limitations (insufficient detail preservation, training instability, poor environmental adaptability). By synthesizing recent relevant studies, it evaluates four representative models: 1D Cycle-GAN and SN_1D CycleGAN for bearing fault diagnosis (1D vibration signals), and AG-CycleGAN and Enhanced-CycleGAN for steel/IC surface defect detection (e.g., scratch, crack, and pit defects). Analysis shows scenario- specific designs drive efficacy: simulation-driven transfer learning (1D Cycle-GAN) mitigates bearing data scarcity, spectral normalization (SN_1D CycleGAN) stabilizes training dynamics, attention mechanisms (AG- CycleGAN) improve defect contrast, and multi-scale mechanisms (Enhanced-CycleGAN) address scale variability problems. The review concludes industrial CycleGAN success depends on aligning innovations with scenario pain points, establishing a practical framework to guide industrial-oriented model design.